Statistical learning theory

Материал из Wiki - Факультет компьютерных наук
Перейти к: навигация, поиск


- All lecture notes, seminars and solutions in 1 pdf.

- Marks for the first module.

- Solutions of the seminar of 12 December.

- The teacher for the second module is Quentin Paris. You can find all materials here.

Mistakes and Russian texts

Some number of students have trouble with math and English. Email questions! Mistakes, questions and answers are here.

Russian texts: Tatiana has send me the following links that might help those who have trouble with English. A lecture on VC-dimensions was given by K. Vorontsov. A course on Statistical Learning Theory by Nikita Zhivotovsky is given at MIPT. Some short description about PAC learning on p136 in the book ``Наука и искусство построения алгоритмов, которые извлекают знания из данных, Петер Флах. On you can find brief and clear definitions.

Exams module 1

Consultation: Monday 30th of Oktober, 9h30-11h50 classroom 435: I will be answering questions to all interested students.

There are two exams.

Problems exam: Tuesday 31 Okt. 12h10-15h00: The score of this exam has weight 0.2 in your final grade. You solve exercises similar to the ones in the seminars. You can bring lecture notes, solutions of the problem lists, handwritten notes, and pages from Chapt 3, Sect. 4.4 and Chapt 6 from the book "Foundations of Machine Learning Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar".

Colloquium exam: This exam counts for 0.2 of your final grade. You will receive a lemma, proposition or theorem from the lecture notes (and a few topics from the seminars). You need to write the proof and the teacher will ask questions to check your understanding. You can bring a single A4 sheet with things that you find hard to remember. You can write on both sides of this sheet. You can know your subgroup from this list.

List of questions for the colloquium.

Group Date Time Room
БПМИ 141-1 Wednesday 1st of November 12h10-15h40 219
БПМИ 141-2 Wednesday 1st of November 13h40-16h10 219
БПМИ 142-1 Wednesday 1st of November 16h40-18h40 219
БПМИ 142-2 Wednesday 1st of November 17h40-19h40 219
БПМИ 143+145 Thursday 2th of November 15h10-17h10 219
БПМИ 144 Thursday 2th of November 16h40-18h40 219
3th year Friday 3th of November 15h10-17h40 219

Your score of the homework has weight 0.1 in your final grade. Activities in the second module count for 0.5 of weight to the final grade.


Homework module 1

General Information

Syllabus for the 1st module

Course materials

Date Summary Lecture notes Problem list Solutions
5 sept PAC-learning and VC-dimension: definitions 1st and 2nd lecture Updated on 13th of Sept. Problem list 1 Solutions list 1
12 sept PAC-learning and VC-dimension: proof of fundamental theorem Problem list 2 Solutions list 2
19 sept Sauer's lemma, neural networks and agnostic PAC-learning 3th lecture Updated on the 23th of Sept. Problem list 3 Solutions list 3
26 sept Measure concentration, agnostic PAC-learning and Computational learning theory 4th lecture Problem list 4 Solutions list 4
3 okt Agnostic learning and the adaBoost algorithm 5th lecture (21st of Okt. added part about comp. learning and Boosting) Problem list 5
10 okt Boosting: risk bounds using Rademacher complexities 6th lecture (Update 27th of Okt.) Mohri's book: p33-40 Problem list 6 See lecture notes.
17 okt Margin theory and a deep boosting algorithm Mohri's book: p75-83, p131-136 Problem list 7 Solutions list 7

For the last lecture: on the exams there will only be questions about the seminar. The materials of the theory lectures will be covered again in more detail in the 2nd module.

A gentle introduction to the materials of the first 3 lectures and an overview of probability theory, can be found in chapters 1-6 and 11-12 of the following book: Sanjeev Kulkarni and Gilbert Harman: An Elementary Introduction to Statistical Learning Theory, 2012.

Foundations of machine learning, Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalker, 2012. These books can downloaded from .

(We will study a new boosting algorithm, based on the paper: Multi-class deep boosting, V. Kuznetsov, M Mohri, and U. Syed, Advances in Neural Information Processing Systems, p2501--2509, 2014. Notes will be provided.)

Office hours

Person Monday Tuesday Wednesday Thursday Friday
Bruno Bauwens 15:05–18:00 15:05–18:00 Room 620
Quentin Paris